Yulia E Tomilova ¹ Nikolay E Russkikh ² Igor M Yi ² Elizaveta V Shaburova ³ Viktor N Tomilov ⁴ Galina B Pyrinova ¹ Svetlana O Brezhneva ¹ Olga S Tikhonyuk ¹ Nadezhda S Gololobova ¹ Dmitriy V Popichenko ¹ Maxim O Arkhipov ¹ Leonid O Bryzgalov ¹ Evgeny V Brenner ¹ Anastasia A Artyukh ¹ Dmitry N Shtokalo ^2,3 Denis V Antonets ^3* Mikhail K Ivanov ^2,5

• ¹ AO Vector-Best, Novosibirsk, Russia
• ² AcademGene LLC, Novosibirsk, Russia
• ³ MSU Institute for Artificial Intelligence, Lomonosov Moscow State University, Moscow, Russia
• ⁴ SibEnzyme Ltd, Novosibirsk, Russia
• ⁵ Institute of Molecular and Cellular Biology SB RAS, Novosibirsk, Russia

Modification of natural enzymes to introduce new properties and enhance existing ones is a central challenge in bioengineering. This study is focused on the development of Taq polymerase mutants that show enhanced reverse transcriptase (RTase) activity while retaining other desirable properties such as fidelity, 5'-3' exonuclease activity, effective deoxyuracyl incorporation, and tolerance to locked nucleic acid (LNA)-containing substrates. Our objective was to use AI-driven rational design combined with multiparametric wet-lab analysis to identify and validate Taq polymerase mutants with an optimal combination of these properties. The experimental procedure was conducted in several stages: 1) 18 candidate mutations across six sites were selected from literature for experimental evaluation along with the wild type enzyme to make the initial training dataset; 2) using Taq polymerase variants embeddings obtained from protein language model we trained a Ridge regression model to predict multiple enzyme properties to select 14 more candidates for experimental evaluation, expanding the dataset for further refinement; 3) to better manage the risk by assessing predictions confidence intervals we transitioned to Gaussian process regression with our expanded dataset comprising 33 data points; 4) with this enhanced model, we conducted the in silico screening of over 18 million mutations and narrowed the field to 16 top candidates for comprehensive wet-lab evaluation. This iterative, data-driven strategy ultimately led to the identification of 18 enzyme variants that exhibited markedly improved RTase activity while maintaining a favorable balance of other key properties. These enhancements were generally accompanied by lower Kd, moderately reduced fidelity, and greater tolerance to noncanonical substrates, thereby illustrating a strong interdependence among these traits. Several enzymes validated via this procedure were effective in single-enzyme real-time reverse-transcription PCR setups, implying their utility for the development of new tools for real-time reverse-transcription PCR technologies, such as pathogen RNA detection and gene expression analysis. This study illustrates how AI can be effectively integrated with experimental bioengineering to enhance enzyme functionality systematically. Our approach offers a robust framework for designing enzyme mutants tailored to specific biotechnological applications. The results of our biological activity predictions for mutated Taq polymerases can be accessed at https://huggingface.co/datasets/nerusskikh/taqpol_insilico_dms.